def __init__(self, state_size, action_size, seed, memory_type):
# keep params
self.state_size = state_size
self.action_size = action_size
self.seed = random.seed(seed)
# Create local and target network
self.qnn_local = QNetwork(state_size,
action_size,
seed,
fc1_units=74,
fc2_units=74).to(device)
self.qnn_target = QNetwork(state_size,
action_size,
seed,
fc1_units=74,
fc2_units=74).to(device)
# Init optimizer
self.optimizer = optim.Adam(self.qnn_local.parameters(), lr=LR)
# Replay Memory
if memory_type == 0:
print("Using FifoMemory...")
self.memory = FifoMemory(BUFFER_SIZE)
else:
print("Using simple DequeMemory...")
self.memory = DequeMemory(action_size, BUFFER_SIZE, BATCH_SIZE,
seed)
self.t_step = 0
def __init__(self, state_size, action_size, random_seed):
"""Initialize an Agent object.
Params
======
state_size (int): dimension of each state
action_size (int): dimension of each action
random_seed (int): random seed
"""
# Store parameters
self.state_size = state_size
self.action_size = action_size
self.seed = random.seed(random_seed)
self.epsilon = EPSILON
# Actor Network (w/ Target Network)
self.actor_local = Actor(state_size, action_size,
random_seed).to(device)
self.actor_target = Actor(state_size, action_size,
random_seed).to(device)
self.actor_optimizer = optim.Adam(self.actor_local.parameters(),
lr=LR_ACTOR)
# Critic Network (w/ Target Network)
self.critic_local = Critic(state_size, action_size,
random_seed).to(device)
self.critic_target = Critic(state_size, action_size,
random_seed).to(device)
self.critic_optimizer = optim.Adam(self.critic_local.parameters(),
lr=LR_CRITIC,
weight_decay=WEIGHT_DECAY)
# Noise process
self.noise = OUNoise(action_size, random_seed)
# Replay memory
self.memory = FifoMemory(BUFFER_SIZE, BATCH_SIZE)
# Short term memory contains only 1/100 of the complete memory and the most recent samples
self.memory_short = FifoMemory(int(BUFFER_SIZE / 100), int(BATCH_SIZE))
def __init__(self, state_size, action_size, random_seed=1):
"""Initialize an Agent object.
Params
======
state_size (int): dimension of each state
action_size (int): dimension of each action
random_seed (int): random seed
"""
# Store parameters
self.state_size = state_size
self.action_size = action_size
self.seed = random.seed(random_seed)
self.epsilon = EPSILON
self.lr_actor = LR_ACTOR
self.lr_critic = LR_CRITIC
self.lr_decay = WEIGHT_DECAY
# Actor Network (w/ Target Network)
self.actor_local = Actor(state_size, action_size, random_seed).to(device)
self.actor_target = Actor(state_size, action_size, random_seed).to(device)
self.actor_optimizer = optim.Adam(self.actor_local.parameters(), lr=self.lr_actor)
# Critic Network (w/ Target Network)
self.critic_local = Critic(state_size, action_size, random_seed).to(device)
self.critic_target = Critic(state_size, action_size, random_seed).to(device)
self.critic_optimizer = optim.Adam(self.critic_local.parameters(), lr=self.lr_critic)
# Noise process
self.noise = OUNoise(action_size, random_seed)
# Replay memory
self.memory = FifoMemory(BUFFER_SIZE, BATCH_SIZE)
# Success memory contains only the last 10 samples which led to a positive reward
self.memory_success = FifoMemory(int(BUFFER_SIZE), int(BATCH_SIZE))
# Rolling sample memory of last 10 samples
self.memory_short = FifoMemory(10, 10)
class AgentTD3():
"""Interacts with and learns from the environment."""
def __init__(self, state_size, action_size, random_seed):
"""Initialize an Agent object.
Params
======
state_size (int): dimension of each state
action_size (int): dimension of each action
random_seed (int): random seed
"""
# Store parameters
self.state_size = state_size
self.action_size = action_size
self.seed = random.seed(random_seed)
self.epsilon = EPSILON
# Actor Network (w/ Target Network)
self.actor_local = Actor(state_size, action_size,
random_seed).to(device)
self.actor_target = Actor(state_size, action_size,
random_seed).to(device)
self.actor_optimizer = optim.Adam(self.actor_local.parameters(),
lr=LR_ACTOR)
# Critic Network (w/ Target Network)
self.critic_local = Critic(state_size, action_size,
random_seed).to(device)
self.critic_target = Critic(state_size, action_size,
random_seed).to(device)
self.critic_optimizer = optim.Adam(self.critic_local.parameters(),
lr=LR_CRITIC,
weight_decay=WEIGHT_DECAY)
# Noise process
self.noise = OUNoise(action_size, random_seed)
# Replay memory
self.memory = FifoMemory(BUFFER_SIZE, BATCH_SIZE)
# Short term memory contains only 1/100 of the complete memory and the most recent samples
self.memory_short = FifoMemory(int(BUFFER_SIZE / 100), int(BATCH_SIZE))
def step(self, state, action, reward, next_state, done, timestep):
"""Save experience in replay memory, and use random sample from buffer to learn."""
# Save experience / reward
self.memory.add(state, action, reward, next_state, done)
self.memory_short.add(state, action, reward, next_state, done)
# Learn at defined interval, if enough samples are available in memory
# HINT from Udacity "benchmark": learn every 20 timesteps and train 10 samples
if len(self.memory) > BATCH_SIZE and timestep % LEARN_EVERY == 0:
for _ in range(LEARN_NUM):
experiences = self.memory.sample()
experiences_short = self.memory_short.sample()
# delay update of the policy and only update every 2nd training
self.learn(experiences_short, timestep % 2, GAMMA)
self.learn(experiences, timestep % 2, GAMMA)
def act(self, state, add_noise=True):
"""Returns actions for given state as per current policy."""
state = torch.from_numpy(state).float().to(device)
self.actor_local.eval()
with torch.no_grad():
action = self.actor_local(state).cpu().data.numpy()
self.actor_local.train()
# TD3 --> Action noise regularisation
if add_noise:
action += self.epsilon * self.noise.sample()
# The range of noise is clipped in order to keep the target value
# close to the original action.
clipped_action = np.clip(action, -1, 1)
return clipped_action
def reset(self):
self.noise.reset()
def learn(self, experiences, delay, gamma):
"""Update policy and value parameters using given batch of experience tuples.
Q_targets = r + γ * critic_target(next_state, actor_target(next_state))
where:
actor_target(state) -> action
critic_target(state, action) -> Q-value
Params
======
experiences (Tuple[torch.Tensor]): tuple of (s, a, r, s', done) tuples
gamma (float): discount factor
"""
states, actions, rewards, next_states, dones = experiences
# ---------------------------- update critic ---------------------------- #
# TD3 --> Using a pair of critic networks (The twin part of the title)
# Get predicted next-state actions and Q values from target models
actions_next = self.actor_target(next_states)
Q_targets_next1, Q_targets_next2 = self.critic_target(
next_states, actions_next)
# TD3 --> Take the minimum of both critic in order to avoid overestimation
Q_targets_next = torch.min(Q_targets_next1, Q_targets_next2)
# Compute Q targets for current states (y_i)
Q_targets = rewards + (gamma * Q_targets_next * (1 - dones))
# Compute critic loss
Q_expected1, Q_expected2 = self.critic_local(states, actions)
# compute critic loss [HOW MUCH OFF?] as sum of both loss from target
critic_loss = F.mse_loss(Q_expected1, Q_targets) + F.mse_loss(
Q_expected2, Q_targets)
# minimize loss [TRAIN]
self.critic_optimizer.zero_grad()
critic_loss.backward()
torch.nn.utils.clip_grad_norm_(self.critic_local.parameters(), 1)
self.critic_optimizer.step()
# ---------------------------- update actor ---------------------------- #
# TD3 --> Delayed updates of the actor = policy (The delayed part)
# Compute actor loss
if delay == 0:
actions_pred = self.actor_local(states)
# compute loss [HOW MUCH OFF?]
actor_loss = -self.critic_local.Q1(states, actions_pred).mean()
# minimize loss [TRAIN]
self.actor_optimizer.zero_grad()
actor_loss.backward()
self.actor_optimizer.step()
# ----------------------- update target networks ----------------------- #
self.soft_update(self.critic_local, self.critic_target, TAU)
self.soft_update(self.actor_local, self.actor_target, TAU)
# ---------------------------- update noise ---------------------------- #
self.epsilon -= EPSILON_DECAY
self.noise.reset()
def soft_update(self, local_model, target_model, tau):
"""Soft update model parameters.
θ_target = τ*θ_local + (1 - τ)*θ_target
Params
======
local_model: PyTorch model (weights will be copied from)
target_model: PyTorch model (weights will be copied to)
tau (float): interpolation parameter
"""
for target_param, local_param in zip(target_model.parameters(),
local_model.parameters()):
target_param.data.copy_(tau * local_param.data +
(1.0 - tau) * target_param.data)
class TD3Agent():
"""Interacts with and learns from the environment."""
def __init__(self, state_size, action_size, random_seed=1):
"""Initialize an Agent object.
Params
======
state_size (int): dimension of each state
action_size (int): dimension of each action
random_seed (int): random seed
"""
# Store parameters
self.state_size = state_size
self.action_size = action_size
self.seed = random.seed(random_seed)
self.epsilon = EPSILON
self.lr_actor = LR_ACTOR
self.lr_critic = LR_CRITIC
self.lr_decay = WEIGHT_DECAY
# Actor Network (w/ Target Network)
self.actor_local = Actor(state_size, action_size, random_seed).to(device)
self.actor_target = Actor(state_size, action_size, random_seed).to(device)
self.actor_optimizer = optim.Adam(self.actor_local.parameters(), lr=self.lr_actor)
# Critic Network (w/ Target Network)
self.critic_local = Critic(state_size, action_size, random_seed).to(device)
self.critic_target = Critic(state_size, action_size, random_seed).to(device)
self.critic_optimizer = optim.Adam(self.critic_local.parameters(), lr=self.lr_critic)
# Noise process
self.noise = OUNoise(action_size, random_seed)
# Replay memory
self.memory = FifoMemory(BUFFER_SIZE, BATCH_SIZE)
# Success memory contains only the last 10 samples which led to a positive reward
self.memory_success = FifoMemory(int(BUFFER_SIZE), int(BATCH_SIZE))
# Rolling sample memory of last 10 samples
self.memory_short = FifoMemory(10, 10)
def update_model(self, state, action, reward, next_state, done):
self.step(state, action, reward, next_state, done)
def step(self, state, action, reward, next_state, done):
"""Save experience in replay memory, and use random sample from buffer to learn."""
reached = True
if len(self.memory_success) < BATCH_SIZE:
reached = False
self.memory.add(state, action, reward, next_state, done)
self.memory_short.add(state, action, reward, next_state, done)
# Fill the success memory in case this agents receives positive reward
if reward > 0.0:
for i in range(len(self.memory_short)):
self.memory_success.add(
self.memory_short.samples[i].state, \
self.memory_short.samples[i].action, \
self.memory_short.samples[i].reward, \
self.memory_short.samples[i].next_state, \
self.memory_short.samples[i].done)
self.memory_short.clear()
if reached == False and len(self.memory_success) > BATCH_SIZE:
print("Success memory ready for use!")
# Train with the complete replay memory
if len(self.memory) > BATCH_SIZE:
for i in range(LEARN_NUM_MEMORY):
experiences = self.memory.sample()
# delay update of the policy and only update every 2nd training
self.learn(experiences, 0 , GAMMA)
# Train with the success replay memory
if (len(self.memory_success) > self.memory_success.batch_size):
for i in range(LEARN_NUM_MEMORY_SUCCESS):
experiences_success = self.memory_success.sample()
self.learn(experiences_success, 0 ,GAMMA)
def act(self, state, add_noise=True):
"""Returns actions for given state as per current policy."""
state = torch.from_numpy(state).float().to(device)
self.actor_local.eval()
with torch.no_grad():
action = self.actor_local(state).cpu().data.numpy()
self.actor_local.train()
# TD3 --> Action noise regularisation
if add_noise:
action += self.epsilon * self.noise.sample()
# The range of noise is clipped in order to keep the target value
# close to the original action.
clipped_action = np.clip(action, -1, 1)
self.epsilon *= EPSILON_DECAY
return clipped_action
def reset(self):
self.noise.reset()
def learn(self, experiences, delay ,gamma):
"""Update policy and value parameters using given batch of experience tuples.
Q_targets = r + γ * critic_target(next_state, actor_target(next_state))
where:
actor_target(state) -> action
critic_target(state, action) -> Q-value
Params
======
experiences (Tuple[torch.Tensor]): tuple of (s, a, r, s', done) tuples
gamma (float): discount factor
"""
states, actions, rewards, next_states, dones = experiences
# ---------------------------- update critic ---------------------------- #
# TD3 --> Using a pair of critic networks (The twin part of the title)
# Get predicted next-state actions and Q values from target models
actions_next = self.actor_target(next_states)
Q_targets_next1, Q_targets_next2 = self.critic_target(next_states, actions_next)
# TD3 --> Take the minimum of both critic in order to avoid overestimation
#Q_targets_next = torch.min(Q_targets_next1, Q_targets_next2)
Q_targets_next = Q_targets_next1
# Compute Q targets for current states (y_i)
Q_targets = rewards + (gamma * Q_targets_next * (1 - dones))
# Compute critic loss
Q_expected1, Q_expected2 = self.critic_local(states, actions)
# compute critic loss [HOW MUCH OFF?] as sum of both loss from target
#critic_loss = F.mse_loss(Q_expected1, Q_targets)+F.mse_loss(Q_expected2, Q_targets)
critic_loss = F.mse_loss(Q_expected1, Q_targets)
# minimize loss [TRAIN]
self.critic_optimizer.zero_grad()
critic_loss.backward()
torch.nn.utils.clip_grad_norm_(self.critic_local.parameters(), 1)
self.critic_optimizer.step()
# ---------------------------- update actor ---------------------------- #
# TD3 --> Delayed updates of the actor = policy (The delayed part)
# Compute actor loss
if delay == 0:
actions_pred = self.actor_local(states)
# compute loss [HOW MUCH OFF?]
actor_loss = -self.critic_local.Q1(states, actions_pred).mean()
# minimize loss [TRAIN]
self.actor_optimizer.zero_grad()
actor_loss.backward()
self.actor_optimizer.step()
# ----------------------- update target networks ----------------------- #
self.soft_update(self.critic_local, self.critic_target, TAU)
self.soft_update(self.actor_local, self.actor_target, TAU)
# ---------------------------- update noise ---------------------------- #
self.noise.reset()
def soft_update(self, local_model, target_model, tau):
"""Soft update model parameters.
θ_target = τ*θ_local + (1 - τ)*θ_target
Params
======
local_model: PyTorch model (weights will be copied from)
target_model: PyTorch model (weights will be copied to)
tau (float): interpolation parameter
"""
for target_param, local_param in zip(target_model.parameters(), local_model.parameters()):
target_param.data.copy_(tau*local_param.data + (1.0-tau)*target_param.data)